Datasets:

carlosdanielhernandezmena

/

ravnursson_asr

like 2

Hello Haukur,

Thanks for your nice questions.

• "Are all dialects represented in the dev/test splits? The description does not say".

No, the main criteria to create the Dev and Test splits was the gender balancing of the speakers.

• "I assume that prompts mean normalized_text as they are called in the dataset.":

Yes, that is correct.

• " It does not say whether the prompts in the dev/test sets are present in the training set."

Most the sentences in the corpus are repeated at least once by another speaker. So, it is almost impossible to create a Dev or Test portion with unique sentences. One possible solution could be to select senteces with exactly n repetitions and put them in the DEV and Test portions. But doing so, the problem now would be that all portions would share spekaers. So, we had to make a decision, which was: not to share speakers.

• " Having the same prompts in the training set and then also in the dev/test causes a serious data leak as a model might start guessing the whole prompt when it hears something which sounds alike".

I am not completely agree with this. Every speaker is different and due to the concept of "speech variability" it is possible for the model to learn acoustic features from the speakers that are present in the training set and at the same time to ignore acoustic features present in the speakers of the Dev and Test portions.

On the other hand, we aware users about the problems of using the training prompts in a language modeling task in the section "Discussion of Biases" of the dataset card of the corpus.

Hope this helps,
Carlos